feat(worker): skip-on-error no loop + conflict auto-fix + deduped notif

1. Skip ONCE on infra errors (Claude Code CLI missing/timeout/unreachable):
   - run_ai_fix returns '[INFRA] ...' reasons; worker marks the PR permanently
     skipped at that head SHA in fix-state (no more retry every 5 min)
   - skip is recorded per {repo,pr,sha}; cleared when head SHA changes
2. Merge conflict auto-fix via Claude Code:
   - mergeable=False or merge HTTP 409 now trigger run_ai_fix (prompt already
     merges base + resolves conflicts) once per head SHA
   - success → next tick re-checks mergeable and merges; failure → skip once
3. Discord skip notification dedupe:
   - notify_skip_once(): posts '⏭️ Skipped: <reason>' exactly once per
     PR+head_sha (state.notified flag); no repeated spam every cron tick
   - skip reason + which PR is visible in the notification
4. State migration: legacy {repo:{pr:'sha'}} → dict form handled in _pr_entry
Verified: py_compile clean, state-helper unit tests pass (skip/fixed/notify
dedupe/legacy migration). Cron wrapper execs repo copy — no manual sync.

.hermes/plans/2026-09-21-queue-worker-resiliency.md

@@ -0,0 +1,67 @@
+# Queue Worker Resiliency — Skip on Error + Conflict Auto-fix + Deduped Notify
+
+## Goal
+Make pr-queue-worker.py resilient: no infinite looping on infra errors,
+auto-resolve merge conflicts via Claude Code, and notify Discord about skips
+exactly once (no repeated spam).
+
+## Scope
+- `scripts/pr-queue-worker.py` only (cron worker, canonical in repo)
+- State persisted in `/tmp/pr-queue-fix-state.json` (already exists)
+
+## Changes
+
+### A. Skip-on-error (no loop) — fix-state gains skip reasons
+Current `FIX_STATE_FILE` shape: `{repo: {pr: head_sha}}` (fixed map).
+Extend value to either a string SHA (fixed) or a dict for skip:
+  `{repo: {pr: {"fixed_sha": "<sha>", "skip_reason": "<reason>", "skip_sha": "<sha>"}}}`
+- `run_ai_fix` failure kinds:
+  - CLI not found / not executable → **infra error** → mark skipped, notify once
+  - timeout → mark skipped (Claude hung), notify once
+  - "Claude ran but no push" → NOT infra error (Claude worked, nothing to fix)
+    → allow retry next tick (cheap) OR mark fixed? — decide: keep retry (this
+    is often a transient "no changes needed" state, but re-running every tick
+    is exactly the loop user complains about) → **mark skipped too** (reason
+    "no changes produced"), notify once, so it stops looping. User can re-open.
+  - clone failed → transient (network) → retry next tick (NOT skip) — but cap:
+    track clone-fail attempts, after 3 → skip w/ notify.
+- `already_fixed` semantics: returns True also for skip state (same SHA) so we
+  don't re-run OR re-notify.
+- New helper `mark_skip(state, repo, pr, sha, reason)` + `get_skip(state,...)`.
+
+### B. Merge conflict → Claude Code resolve
+- When `mergeable is False` (GitHub says conflict) OR merge returns 409:
+  - If not already attempted at this SHA (fix-state): run `run_ai_fix` (its
+    prompt already includes "merge base + resolve conflicts intelligently").
+  - If fix returns True → continue to re-check mergeability next iteration
+    (head SHA changed; loop naturally proceeds).
+  - If fix returns False (error/no push) → mark skip + notify once.
+- Prevent infinite conflict-fix loop: one AI attempt per head SHA (fix-state
+  `fixed_sha` covers it). If Claude pushes but conflict persists → next tick
+  `mergeable is False` again but fix-state SHA = old → we'd re-run. Fix:
+  mark conflict-resolve-attempted in state with the head SHA and the *result*
+  ("conflict-fix-pushed" / "conflict-fix-failed"); skip re-run on same SHA.
+
+### C. Deduped Discord notification for skips
+- State file gains `notified: {repo: {pr: {"reason": sha}}}` (or embed in skip
+  entry: when we notify, record sha; skip notify if same sha+reason already
+  notified).
+- Notification: one embed per skip reason, listing PRs:
+  `⚠️ Skipped: #19 (dependabot) — Clode Code CLI not found` — only on first
+  occurrence per PR+reason, not every 5-min tick.
+- The full run report (flush_log) still goes out each tick — that's the
+  periodic status, fine. The dedupe applies to *skip-specific* notifications.
+
+## Files touched
+- `scripts/pr-queue-worker.py`
+
+## Verification
+- Unit-ish: run worker with a fake PR? Hard (needs GitHub). Instead:
+  - py_compile
+  - dry-run logic test: monkeypatch run_ai_fix to return False with
+    "Claude Code CLI not found" → assert mark_skip called once, second run
+    skips re-attempt, notify once (check state file).
+  - conflict path: monkeypatch mergeable False + run_ai_fix True → assert
+    conflict-fix-attempt recorded once.
+- Live: after deploy, open a tiny conflicting PR (or use an existing one) and
+  watch one skip notification, then verify no repeat next tick.